They refer to this in the paper as a part of the "cold start data" which they use to fine-tune DeepSeek-V3 prior to training R1. They don't specifically name O

They refer to this in the paper as a part of the "cold start data" which they use to fine-tune DeepSeek-V3 prior to training R1. They don't specifically name OpenAI, but they refer to "directly prompting models to generate answers with reflection and verification".